Is the latest v4 flash model from deepseek, with 284B parameters, performing close to GPT-5.6 Sol? Pretty impressive, right? But what if a 35B model could achieve similar performance? Tsinghua's latest paper has achieved this seemingly impossible goal using recursive self-improvement (RSI)! Paper title: Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering Key points: True recursive self-improvement (RSI) requires AI systems to improve "the process of building AI itself" (aka AI4AI). Machine Learning Engineering (MLE) provides a concrete, executable, and verifiable testing platform since code can actually run and provide execution feedback. They’ve released a fully open-source full-stack system called OpenMLE, specifically designed for this type of research, featuring three main components: • OpenMLE-Gym: Verifiable task environments + execution feedback • OpenMLE-RL / ERL: Operator learning (execution-based SFT + RL) • OpenMLE-Evo: Long-term search (experience-guided evolutionary search) Building on this, they trained a 35B parameter meta-evolution agent called Frontis-MA1. This model aligns around four atomic-level program evolution operators: 1. Draft (drafting initial solutions) 2. Improve (refining based on scores and execution evidence) 3. Debug (fixing faulty code) 4. Crossover (combining strengths of two parent solutions) These operators are trained via execution-grounded SFT and RL, then combined during inference into long-term search loops, enabling "learning" and "evolution" to run within the same closed loop. Key results (under strict resource constraints): On MLE-Bench Lite, with only a 12-hour budget per task and a single RTX 4090 GPU (12GB VRAM limit): • Base model: Medal Average 39.39% • With OpenMLE-Evo: 60.61% • Using OpenMLE-Evo-Max (asynchronous search + benchmark-independent experience priors): 71.21% This performance already surpasses GPT-5.5 + Codex and approaches GPT-5.6 Sol and the 2.8T parameter Kimi K3. It also shows transferability on the held-out test set NatureBench Lite: • Fixed framework, swapping in the trained model: Match-SOTA improved from 50% → 70% • Fixed model, swapping in the search framework: improved from 20% → 50% This demonstrates that both model fine-tuning and the search framework contribute gains and are transferable. Significance and highlights: This isn’t just another RSI study—it brings RSI to a level that’s reproducible and executable on consumer-grade hardware. RSI is no longer the exclusive domain of big corporations. Now, even individuals with a 4090 GPU can continuously iterate and improve their models on their own.